Meta Reportedly Saved $3.9 Billion in Taxes by Treating AI Infrastructure as Research

Meta saved a reported $3.9 billion through the federal research tax credit in 2025, making it the largest beneficiary among publicly traded companies. The Decoder, citing new New York Times reporting, describes a strategy that treats AI data centers as “pilot models” and Nvidia chips as experimental materials—a research classification with billion-dollar consequences for Meta’s tax bill.

The September 30 disclosure puts the tax treatment of AI infrastructure at the center of a dispute over what counts as research. Meta defends its use of the credit by pointing to $200 billion in research and development spending over five years. Its own securities filings, however, warn that the tax-credit claims could be challenged.

The experimental label behind the savings

The mechanism is the classification of physical infrastructure as experimental research assets. According to the Times reporting summarized by The Decoder, Meta uses two labels to support its claims:

• AI data centers are classified as “pilot models.”
• Nvidia chips are classified as experimental materials.

That framing sits alongside a commercial purpose Meta has described publicly. In January 2025, Mark Zuckerberg said its data centers would “drive our core products and business.” In July of that year, he outlined plans to invest hundreds of billions of dollars in computing infrastructure to build superintelligence, anchored by several multi-gigawatt clusters.

The credit’s original sponsor questions how far that use has stretched. The provision dates to a 1981 law introduced by then-congressman James Shannon. He told the Times it was intended for “people power, knowledge, information,” and said Meta’s use had “gone way, way beyond what anybody could have imagined.” That is a criticism of the credit’s reach, not a legal ruling on Meta’s claims.

Meta’s defense—and its auditor’s role

Meta’s response emphasizes the scale of its research work: the company says it spent $200 billion on R&D over the preceding five years. That defense addresses its overall commitment to research. The disputed classification concerns the equipment and facilities used to carry out its AI activities.

EY, Meta’s auditor, reportedly approved the approach and helped establish it. The Times also reported that EY is pitching the same approach to other companies seeking to offset AI chip purchases. Its role therefore extends beyond reviewing Meta’s accounts: the firm helped develop a strategy it is now offering to other buyers of AI hardware.

Approval by its auditor sits alongside a warning in Meta’s SEC filings that the savings could face challenges. The company’s reserves for uncertain tax positions rose 45% to $18.74 billion, The Decoder reports. Those reserves concern uncertain tax positions; they should not be read as a stated repayment amount for the AI research credits.

Research incentive or equipment subsidy?

In commentary published the same day, the Institute on Taxation and Economic Policy argues that the research credit should encourage experimentation, not subsidize equipment a technology company needs to operate. It cites the same Times reporting, rather than providing independent corroboration of Meta’s infrastructure classifications.

ITEP also questions whether an incentive changes the spending decisions of a company with ample cash. It points to Meta’s $43 billion in cash and cash equivalents at the beginning of 2025 and argues that the company already had the means to fund its plans. That is a policy argument about the value of the subsidy, distinct from whether particular expenses legally qualify.